CONFIDENTIAL · PRE-SEED · 2026
Autonomous software engineering, self-hosted

Coding agents work. The token bill is what breaks them.

CoreForge Atlas runs the full plan–write–test–ship loop on infrastructure we operate ourselves, instead of metering every task through someone else's frontier API.

The raiseWhere we actually are
OR EMAIL sarmed@creatorcoreforge.com
01Problem

Every agent company is renting its margin.

Autonomous coding agents have crossed the line from demo to useful. The category leaders — Devin, Cursor Agent, Copilot Workspace, Cognition — all run on frontier APIs. A single agent task consumes an enormous number of tokens, because the loop retries, reads, tests, and revises before it produces anything a human sees.

That works while the compute is subsidised by venture funding. It does not survive contact with a real gross-margin conversation. The cost of a task is set by someone else's price list, and it is the largest line item in the business.

02Approach

Own the inference. Rent nothing per task.

Open-weight models have closed most of the capability gap on coding work. Atlas runs a fine-tuned open-weight model on hardware we control, which converts a per-token variable cost into a fixed hourly one. Above a modest utilisation threshold, the economics invert.

We are explicit about what is ours and what is not. The base model is open-weight and Apache 2.0 licensed. The orchestration layer — the sandbox, the verification loop, the workforce system, repository context, integrations, and the Forge self-improvement engine — is the product, and it is where our engineering goes.

03Unit economics

The whole thesis, in one chart.

Effective cost per million tokens of agent work. Self-hosted cost is fixed hourly spend divided by realised throughput — it falls as utilisation rises, while API cost never moves.

Cost per 1M tokens (illustrative)internal model · verify pre-share
Frontier API (GPT-4-class)~$15.00
Self-hosted — low utilisation (~20%)~$8.10
Self-hosted — target utilisation (~75%)~$2.55
Estimates based on A100/H100 spot pricing and observed throughput on a 27B–70B parameter open-weight coding model. Full arithmetic, GPU choice, and throughput measurements available under NDA.
04How it works

A loop, not a chat box.

The difference between generating code and shipping it is verification. Atlas runs each task in an isolated sandbox with the repository's real build and test commands, and does not surface work a human has to babysit.

INGEST

Issue, PR, or natural-language brief. Repo context assembled from git index.

PLAN

Task decomposed into file-level changes by the workforce orchestrator.

EXECUTE

Edits applied in an isolated Docker sandbox with the real toolchain.

VERIFY

Project's own build, test, and lint commands run against the change.

SHIP

Pull request opened for human review through the Approval Center.

05Status

Where we actually are.

Stated plainly, because it will come out in diligence anyway. This is a working product at coreforgeatlas.com — not a slide deck.

ComponentStateNote
Application shellLiveTanStack Start on Cloudflare Workers, auth, RLS, dashboard
Workforce orchestratorLiveMulti-agent hiring, task graph, approval center, audit log
Sandbox executionLiveDocker-isolated per-task env; code editor, terminal, build console
Repository integrationLiveGitHub OAuth + PAT, repo indexing, PR flow
Inference modelLiveOpen-weight 27B running in production; Atlas-350M/2B pretraining on Kaggle for owned stack
Verification loopLiveBuild, test, and lint runners wired end-to-end; hardening reliability at scale
Forge self-improvementIn progressAutonomous suggester → patcher → self-review pipeline
Domain fine-tuneNextPipeline built (Kaggle + RunPod); collecting real trace data
Paid pilotsNextDesign-partner outreach begins with this round
06Moat

What compounds over time.

Owned inference stack. Every task we run generates a training trace. That trace fine-tunes the next model. Competitors renting frontier APIs cannot capture that data legally or economically.

Verification data. Real build/test outcomes across real repositories is the scarcest asset in agentic coding. Atlas produces it by design.

Workforce + Forge. The orchestration layer improves itself through the Forge loop — Atlas patches Atlas, under human approval. Each cycle raises the floor.

07Market

The category is funded. The cost structure is not solved.

Comparables that have raised on the strength of the category alone:

  • Cognition (Devin) — reported $175M Series A at ~$2B, Mar 2024.
  • Cursor / Anysphere — $105M at $2.6B, Aug 2024; further rounds since.
  • Magic.dev — $320M in 2024 on frontier-scale infra thesis.
  • Poolside — $500M Series B at $3B, Oct 2024.

Their architecture is a thin orchestration layer over rented frontier inference — which means their unit economics and ours diverge as volume grows, in our favour. We are not competing on capability first; we are competing on the cost per shipped pull request.

08Raise

What we're raising, and what it buys.

Creator CoreForge Inc. is raising a $5M seed on a SAFE at a $60M post-money cap. Target close Q1 2026. Minimum check $10K; lead allocation available.

55%

Engineering

Verification loop reliability, sandbox hardening, Forge self-improvement, first two hires (systems + ML infra).

30%

Compute

GPU serving capacity (A100/H100 reserved), fine-tuning runs on collected trace data, benchmark evaluations.

15%

Go-to-market

Design partners, onboarding tooling, first paid pilots, developer relations.

09Milestones

Why this round is priced where it is.

Valuation moves with de-risking, not with time. This is the earliest and lowest-priced entry point. Each milestone below removes a specific risk and repricing follows.

Now · this round

Product shell live, loop unproven

Application, workforce, sandbox, GitHub integration all shipped. Verification loop end-to-end reliability is the gating risk.

Month 3 · Milestone 1

End-to-end task completion

Atlas takes a real repository issue and opens a passing pull request without human intervention on 10 consecutive tasks.

Month 6 · Milestone 2

SWE-bench Verified score published

Public benchmark resolve rate, with cost per task disclosed alongside. Self-hosted cost advantage documented on identical workloads.

Month 9 · Milestone 3

Paid design partners

5 teams running Atlas against production repositories under paid pilots. $10K–$25K MRR floor. Seed round follows.

10Team

Who is building this.

Sarmed — Founder & CEO, Creator CoreForge Inc. Solo technical founder. Built and shipped the current Atlas platform end-to-end: the workforce orchestrator, sandbox, GitHub integration, Forge self-improvement engine, and training pipeline. Prior work in software engineering and AI systems.

Investors at this stage are underwriting the founder above everything else. The first two hires funded by this round are an ML infrastructure engineer and a systems engineer for the sandbox and verification path.

11Contact

Come look at the loop.

Happy to walk through the architecture, the cost model, and the code as it stands. No deck-only meetings — you'll see the product running.

Email sarmed@creatorcoreforge.com

sarmed@creatorcoreforge.com · coreforgeatlas.com · Creator CoreForge Inc.